English | 中文 | 日本語
Awesome Seedance
Verified Seedance 2.5 / 2.0 prompt library. 947 cases checked against their original posts, 495 cross-model retests, 27 reusable templates, and 79 installable AI-video Skills. The case library draws from goodcase.ai's 2021 verified AI cases across video, image, UI and copy. Case data syncs daily; new cases land every day.
More verified AI cases with full prompts → GoodCase.ai
If these sources, templates or retest notes help your work, star this repository and suggest a case or model to retest next.
Contents
🚀 Start Here
New to AI video? Follow these five steps and you will have your own clip. Nothing to install.
| Step | What to do |
|---|
| 1 | Pick a look in Prompt Templates by Category. Choose by the pictures. |
| 2 | Open that template (for example UGC creator review) and copy the block under Copy this. |
| 3 | Replace the [bracketed] parts with your own product, person or scene. |
| 4 | Send it to any AI chat (ChatGPT, Claude, Gemini) along with your reference images. You get back a finished prompt. |
| 5 | Paste that prompt into Seedance (Dreamina / Jimeng) and generate. If it looks off, read the template's pitfalls and re-run. |
Templates or Skills? Templates are distilled from 947 verified cases. Skills can apply those structures or guide other parts of production.
| Prompt templates | Skills |
|---|
| Who it is for | Beginners, and anyone who does not want to install anything | People already working in Claude Code, Codex or another coding agent |
| How you use it | Copy, replace the [brackets], paste into any AI chat | One install command, then just tell your agent what you want |
| What it covers | 27 category templates | Case-backed prompt structures and production workflows across 79 Skills and creator methods |
| What you get | One solid prompt at a time | A prompt or a production plan and QA record, depending on the Skill |
Just browsing? Go to Top 30 by heat or all cases. Want proof that a prompt holds up? See the cross-model retests.
🧩 Prompt Templates by Category
27 templates, one per kind of clip, distilled from 899 verified cases. Pick the look you want, open its template, copy the block, swap the [bracketed] parts for your own subject, and hand it to any AI chat. Browse them all in the template index.
🧱 Structural foundations (3 templates)
Cross-cutting skeletons that almost every other template builds on: how to slice a clip into timed beats, and how to pin an identity across those beats.
📱 Realism and UGC (4 templates)
Templates that buy believability by describing camera flaws, body wear and consumer-grade optics instead of asking for quality.
🛒 Commercial and product (5 templates)
Ad-shaped structures where a product has to survive macro shots, hand contact and a hero frame without deforming.
🎭 Narrative and performance (6 templates)
Templates where the payload is a story beat or a line of dialogue rather than a look.
🎨 Stylized animation (3 templates)
Non-photoreal looks that collapse into generic CG unless the style is specified as measurable parameters plus an exclusion list.
💥 Action, dance and effects (6 templates)
Templates driven by body mechanics, beat placement or a physics set-piece rather than by scene description.
🧰 Skills
A Skill is an installable instruction pack for coding agents (Claude Code, Codex and friends). Template Skills help write case-backed prompts; the production workflow plans assets, generation handoff and video review. 36 Skills are below, plus 42 creator methods that carry one creator's signature style.
79 AI video Skills in all: the 63 on goodcase.ai (21 official + 42 creator methods), plus 16 that only appear in the grid above. Full list of the 63 goodcase.ai video Skills → One page, every install line.
Every install line works with the skills CLI. Skills named seedance-… live in this repo under agents/skills: the library and single-kind template Skills are regenerated from case data, while the production workflow is maintained separately. npx seedance-prompt-library install also drops the library Skill straight into Claude Code and Codex. More Skills across image, coding and writing live on goodcase.ai.
🔥 Top 30 by heat
The 30 hottest cases across all versions. prompt opens the full entry in the gallery, source opens the creator's original post.
🔁 Cross-model retests
As far as we know, this is the first public prompt library that re-runs its video prompts on a second model at scale and publishes the result either way. 483 of the cases here have been re-run (495 runs so far), each with a verdict, a judge score and the generated output. A prompt that only ever worked once, for its author, on one model, is a screenshot; a prompt that survives a re-run is a method.
| Model | Runs | Reproduction rate |
|---|
| MiniMax H3 Max 768p | 247 | 72% |
| Grok Imagine | 237 | 74% |
| MiniMax H3 768p | 11 | 82% |
Verdicts across all runs: ✅ 364 reproduced · ⚠️ 123 degraded · ❌ 8 failed. Runs without a final score show as score n/a. Per-case verdicts, scores and output videos are on each case's goodcase.ai page; the model labels and batch dates are explained under Statistics.
Same prompt, second model. Three examples, including one that did not hold up:
Every run costs real inference money: over US$300 across 495 runs so far, at list price with no discounts, which is what anyone else would pay to reproduce them. We publish the result whether or not it flatters the prompt. Want Kling, Veo, Hailuo or the Seedance 2.5 API added to the retest matrix? Sponsor a batch → or write to [email protected].
🎬 All Prompts
All 947 cases, with full prompts, live in the gallery under docs/ (sharded so GitHub renders every page). Start from the gallery index, or jump to a version:
🌐 Browse on goodcase.ai
This README is an index. The full experience lives at goodcase.ai: search across every case, the heat leaderboard, the stability ranking, per-case retest logs with output videos, and the installable Skills that grow out of the cases. Every entry here links back to its goodcase.ai record.
Why this list
Human-verified against the source. Every prompt here was checked against the creator's original post. Prompts reverse-engineered from the output video only, with no source and no submission, are rejected outright, per goodcase.ai's collection standards (in force since 2026-08-05).
Re-run on a second model. Most cases have been re-generated on another video model, with the verdict, score and output published. See Cross-model retests.
Full provenance on every entry. Author, original post link, publish date, and a heat score, a relative percentile among published cases on the same platform. If it didn't rank, it isn't here.
Ships as an installable Agent Skill. npx seedance-prompt-library install drops a template library straight into Claude Code / Codex so your agent writes Seedance prompts from proven structures, not guesses.
Statistics
| Metric | Value |
|---|
| Seedance cases in this repo | 947 |
| Seedance 2.5 | 645 |
| Seedance 2.0 | 302 |
| Unique authors | 210 |
| Re-run on other models | 483 cases / 495 runs |
| Stability score (measured) | 478 cases / avg 78.0 |
| Last updated | 2026-10-30 |
| goodcase.ai, all categories | 2021 cases / 481 creators |
| goodcase.ai, AI video | 1179 cases |
Retest batch note: Retest runs carry two MiniMax labels on purpose. The 2026-08 batches (runIds video-minimax-h3-768p-20260809-phase1 / -20260811-top-heat) ran on Flova's MiniMax H3 768p; the 2026-09 batches (video-minimax-h3-768p-20260906-phase1 / -20260907-phase1) ran on fal.ai's minimax/h3-max/text-to-video endpoint, which only existed from 2026-09-03. They are different model tiers on different platforms, so the labels are kept distinct rather than merged. The 2026-08 entries show finalScore: null because a later human-review pass overwrote their structured evidence notes with a prose verdict; those rows carry a human-reviewed verdict instead, and the original judge scores remain in the private eval evidence manifest.
Each case is counted once; a case tagged with several Seedance versions counts under the highest one.
How to Contribute
New prompt cases go through goodcase.ai's review pipeline so provenance and heat score stay verifiable: submit at goodcase.ai/submit (collection standards: goodcase.ai/standards). Prefer GitHub? Open a pull request that adds one JSON file under submissions/ following submissions/TEMPLATE.json; a maintainer pushes it through the same review, and it lands in data/ on the next export.
Pull requests are welcome for template fixes and new templates in data/templates-local.json, generator and Skill code under scripts/ and agents/, and corrections to English titles or summaries. README.md, README_zh.md, docs/ and the Skill reference are generated from data/, so please don't hand-edit them: change the source, run npm test && npm run generate, and commit the regenerated files in the same PR. See contributing.md for the submission standard, what gets rejected, and how the generator works. This project follows the code of conduct.
🙏 Acknowledgements
This project's format and Skill-packaging approach were shaped by:
👤 Author
Carl (@卡尔的AI沃茨) started GoodCase and maintains this list. Six years of large-model R&D at ByteDance, Alibaba and Baidu before turning to AI video, then a run of AI shorts that made it onto Chinese TV: I Am the Tile Cat (CCTV), When the Stars Shine: AI Bonus (Hunan TV), I Miss You So Much (Zhejiang TV) and Does Birth Decide Fate? I Refuse (Bilibili), each past a million plays. 500k+ followers across platforms, speaker at WAIC and GDPS, guest lecturer on AI at the Communication University of China.
Every case here is something he or the GoodCase reviewers checked against the original post. Follow along on X @aiwarts or the WeChat official account 卡尔的AI沃茨 (Chinese). Questions, corrections and collaborations: [email protected].
Copyright & Takedown Notice
This repository carries three kinds of material under three different terms.
Code (generator scripts, Agent Skill, tooling) is released under the MIT License, see the LICENSE file. The MIT license covers the code only.
Curation (selection, organization, statistics, template extraction, summaries written by us) is licensed under CC BY 4.0. Reuse it with attribution to awesome-seedance / goodcase.ai.
Prompts and media remain the copyright of their original creators. Prompt text, creator summaries, poster images and video references are quoted from publicly published posts for documentation and study; every entry links back to its original source and to its goodcase.ai record. Nothing here grants a license to the underlying prompt or media beyond what the original post allows.
Takedown process. If you are a rights holder and want an entry removed or corrected, open a GitHub issue with the entry's slug (from its goodcase.ai URL) and the original source link, or contact goodcase.ai directly. Requests are verified against the original post and honored on verification; the entry is removed from data/ and disappears from every generated file on the next regeneration.
Star History
License & Reuse
Code in this repository is open source under the MIT License: use it, modify it, build on it, keep the license notice. Curation is CC BY 4.0; prompts and media stay with their creators. Details above under Copyright & Takedown Notice.
